Devtools spent the week turning AI agents into governable, embeddable infrastructure

The week in devtools

The developer-tools sector spent the week pointing in one direction: turning AI agents into governable, embeddable infrastructure rather than novelty assistants. The two freshest moves make it concrete. GitHub shipped AI-generated security detections onto pull requests, reaching languages and frameworks CodeQL never covered, then followed with per-repository Copilot usage metrics at GA and a configurable code-review agent. Resend added OAuth 2.1 with PKCE, letting other products build authenticated experiences on top of its API instead of merely calling it. Different companies, same instinct: expose the primitives that let agents and teams build on you, and hand admins the controls to govern what gets built.

Underneath the agent story is a quieter platform land-grab. Depot made the boldest structural move of the week, entering private beta with a diskless git server that mirrors GitHub repos and plugs straight into its CI. The pattern repeats across the sector: MCP servers, OAuth surfaces, and scoped tokens are becoming table stakes, and the products shipping them are climbing up the stack toward owning more of the developer's workflow. This was a substantive week, not a blog-fed one, though a good chunk of the sector's feed is still marketing content rather than real changelogs.

Leaders

GitHub led on both volume and direction, with one spark and sixteen improvements in the window. AI security detections on pull requests is the headline, but the through-line is Copilot-as-infrastructure: repository-level usage metrics went GA, the code-review agent gained firewalls, custom runners, and head-branch instructions, and admins got PR archiving. The buying criteria it is optimizing for are visibility and policy control, not model novelty.

Depot posted two sparks and is no longer a point tool. Depot Code, its diskless S3-backed git server, arrived days after Depot Metal moved its CI and Sandboxes onto the company's own bare-metal microVM platform. Owning both the compute underneath and the git layer above lets Depot control performance and lock-in end to end while staying interoperable with GitHub and GitLab.

Resend is quietly becoming a developer platform rather than a send endpoint. OAuth 2.1 with PKCE is the clearest signal of that pivot, arriving alongside a hosted MCP server that exposes the same surface to agents and a GitHub secret scanner that adds operational trust. Distribution is shifting toward marketplaces and integrations rather than direct signup.

Cursor kept pushing its agent off the desktop. The fresh work this week refined its Slack surface — the agent now shares a plan before starting, targets named multi-repo environments, and pulls context across channels — building on an iOS public beta that launches cloud agents in isolated VMs and lets developers merge PRs from a phone. The center of gravity is moving from the editor to an orchestration layer for a fleet of agents.

Rootly extended its AI agent from live incident response into the post-incident workflow. New retrospective templates with customizable AI-blocks assemble a first draft from incident data, Slack threads, and call transcripts, keeping every prompt and source visible to the reviewer. Around that core it shipped operational depth — an on-call widget, Intune-protected mobile access, and Cortex catalog sync — the kind of parity work that lets AI-native positioning coexist with enterprise checkboxes.

Wildcards

ElevenLabs is the sector's clearest example of treating agent configuration like a codebase. While others bolt an agent on, it ships branch merge and rebase previews, a rebase endpoint, and per-branch call metrics — version control applied to voice agents. Its in-window spark, Music v2, runs on the other track entirely: a chunk-based composition model that exposes structure and pacing as addressable pieces, broadening the catalog beyond voice.

Merge is a unified-API integrations company repositioning as agent infrastructure. Its Gateway now ships Build Your Own Router, letting customers define routing logic from public benchmarks or private evals, and makes every model reachable by pointing a coding agent — Cursor, Zed, Claude Code, Codex — at a single Gateway key. Routing policies, cost controls, and observability come with it, opening a genuinely new user segment for a company that used to just normalize CRM and HR APIs.

Themes that compounded

• MCP as a default surface: hosted MCP servers showed up across email, CI, PaaS, and observability this week — a fast-hardening interoperability layer rather than an experiment.
• Governance is the enterprise wedge: per-repo usage metrics, object-level permissions, scoped tokens, and configurable review agents all target buyers who purchase on visibility and control.
• Point tools climbing the stack: CI companies are shipping git hosts and compute platforms, and email APIs are shipping auth — the boundary between "tool" and "platform" keeps dissolving.
• AI threaded through the full lifecycle: the agent is moving past a single prompt box into detection, response, retrospective, and code review as connected surfaces.
• Version control as a design metaphor: branches, rebase, and merge previews are being applied to agent config itself, not just source code.

Watch this week

The next moves are legible from what shipped. Expect Depot to widen Depot Code beyond private beta and migrate more workloads onto Depot Metal, and Resend to converge its OAuth and MCP surfaces into a documented "build on Resend" story. GitHub will likely add more Copilot governance surfaces and continue making the code-review agent configurable, while Cursor pushes cloud- and mobile-agent parity and team-scale admin tooling. The signal to track is whether these agents graduate from suggesting actions to taking scoped ones under policy — the difference between an assistant and the governable infrastructure the whole sector is now building toward.